Charging ahead: Prepaid electricity metering in South

Africa

⇤

B. Kelsey Jack†Grant Smith‡

Abstract

The standard approach to recovering the cost of electricity provision is to bill cus-tomers monthly for past consumption. If unable to pay, customers face disconnection,the utility loses revenue, and the service provision model is undermined. A possiblesolution to this problem is prepaid metering, in which customers buy electricity upfrontand use it until the prepaid amount is consumed. We use data from Cape Town, SouthAfrica to examine the effects of prepaid electricity metering on residential consumptionand returns to the electric utility. Over 4,000 customers on monthly billing were invol-untarily assigned to receive a prepaid electricity meter, with exogenous variation in thetiming of the meter replacement. Electricity use falls by about 13 percent as a resultof the switch, a decrease that persists for the following year. This creates a tradeoff forthe utility: revenue from consumption falls but more of it is recovered on time and ata lower cost. The benefits to the electric utility outweigh the costs, on average, thoughresults are very heterogeneous. Poorer customers and those with a history of delinquentpayment behavior show the greatest improvement in profitability when switched to aprepaid meter. These findings point to an important role for metering technologies inexpanding energy access for the poor.

⇤We thank the City of Cape Town, and the Electricity Department in particular, for their cooperationand collaboration on this project. Kathryn McDermott provided outstanding research assistance with allaspects of the project. We also thank Guthrie Gray-Lobe and Grant Bridgman for their assistance with thedata, and Adaiah Lilenstein, Anna Shickele, and the staff of J-PAL Africa for project implementation. TheInternational Growth Centre and J-PAL’s Urban Services Initiative provided financial support. Audiences atBrown University, Harvard University, Tufts University, University of Michigan, and UC Berkeley’s EnergyInstitute provided helpful comments. All errors are our own.

†Department of Economics, Tufts University‡Department of Economics, University of Cape Town

1

1 Introduction

The use of lights measured from space to proxy for economic activity (Henderson et al.2011) drives home the importance of electricity for development. Much of the global increasein electricity consumption in coming decades is forecast to come from developing countries(Wolfram et al. 2012). However, on the ground, expanding access to electrification introducesnew challenges for service providers and for households. The standard model for recoveringthe costs of electricity provision is a postpaid metering system in which a customer receives abill for consumption over the past month. For poor, liquidity constrained customers, findingresources to pay the bill, which is not only a substantial share of income but also varies frommonth to month, presents a major challenge. If unable to pay, customers face disconnection,the utility loses revenue, and the service provision model is undermined. To avoid this,utilities may ration connections to poor households.1

Prepaid electricity meters offer a technological solution to the non-payment problem.These meters, which operate on a debit basis, are increasingly common in developing coun-ties, particularly in Africa. Market forecasts suggest that the greatest growth in electricitymetering in Sub-Saharan Africa will come from prepaid meters, which will dominate theelectricity metering market in Africa by 2020 (Northeast Group 2014). Prepaid meteringis also expanding into the water sector (Heymans et al. 2014).2 South Africa was an earlyadopter of prepaid electricity metering during the phase of rapid electrification following theend of apartheid, when electricity was rolled out to poor and rural communities (Gaunt 2005;Bekker et al. 2008; Dinkelman 2011; van Heusden 2012).

We partner with the municipal government in Cape Town, South Africa to generatenew evidence on the impact of prepaid metering on electricity use and utility revenue andcost recovery. We observe monthly electricity use and payment behavior over a four anda half year period for 4,246 customers in a mandatory meter replacement project in CapeTown. The timing of meter replacement is randomly ordered across 27 geographic areas,and we follow customers for over a year following the replacement. The combination ofinvoluntary replacement and exogenous timing allows us to recover a clean causal estimate

1In practice, this is often done through connection fees, which impose an upfront cost that may be pro-hibitively high for poor households and small businesses (Golumbeanu and Barnes 2013; Lee et al. 2016b,a).Connection fees typically vary with distance to existing infrastructure, so utilities may also choose not toexpand transmission into areas where the customer base is not expected to be profitable.

2The prepayment model has parallels in other technologies. For example, some argue that the rapid riseof mobile phones in Africa is due in part to their reliance on prepayment, which allowed poor users withunpredictable incomes to control their usage and avoid debt.

2

of the impacts of prepaid metering that avoids selection into metering type. Prepaid meteringaffects numerous aspects of consumption, which precludes a clear accounting of consumerwelfare in our setting. Our focus, instead, is on the returns to prepaid metering relative topostpaid billing for the electric utility, which depend on customer behavior.

We find that customers reduce their electricity use when switched from postpaid monthlybills to prepaid electricity metering by 1.9 to 2.2 kWh per customer per day, or 12 to 15percent. The reduction persists for the year following the switch, and is robust to a numberof alternative specifications. The largest reductions come from high consumers, but alsolow property value customers and those who are frequently delinquent in their postpaidbill payments. For more delinquent customers, the switch to prepaid metering represents aparticularly dramatic transition away from the leniency of a weakly enforced monthly bill.On average, customers pay their postpaid bill around 100 days after consumption, or 40days after payment is due. Mechanically, the prepaid meter forces payments to occur beforeconsumption, and payment arrives an average of 3 months sooner as a result of the switchto prepaid metering.

The estimated decrease in consumption results in a corresponding decrease in the amountowed to the utility of around 7 USD per customer per month, and saves the utility around3 USD in kWh supply costs.3 However, better revenue recovery, lower recovery costs, andpayments that arrives around 3 months sooner, on average, tip the balance in favor of prepaidelectricity meters. Specifically, around 2 percent of bills are never paid on postpaid monthlybilling, so the decrease in what customers owe the utility is partially offset by a greaterlikelihood that the amount owed is eventually recovered. In addition, the utility avoidsthe costs of meter reading and bill preparation, and the high costs of enforcement throughdisconnection. All together, the average net gain to the utility over a seven year periodcovers the fixed costs of the prepaid meter. Ignoring customer consumption responses makesprepaid meters look considerably more profitable.

The outcome of the cost-benefit comparison depends on the type of customer, as well asfeatures of the electric utility’s cost environment. Specifically, we observe relatively greaterreturns from switching lower and poorer users (based on average postpaid consumption andproperty values, respectively) to prepaid meters. In addition, customers that typically paidtheir postpaid bills late, carry outstanding debts or had a history of meter disconnections allgenerate substantially greater returns from the switch to prepaid metering than their morereliable counterparts. The electric utility actually loses money by switching customers who

3All values are reported in 2014 real values, based on an exchange rate of 11.45 ZAR/USD.

3

usually pay their bills on time or carry no outstanding debts. We also compare returns toprepaid metering across different administrative costs and observe that higher marginal elec-tricity costs (holding tariffs constant) increase the returns to prepaid metering, because thereduction in consumption creates less of a loss in settings where the average kWh generatesrelatively smaller net revenue. Higher interest rates also increase the advantage of prepaidmetering over postpaid billing.

Together, these results indicate that prepaid electricity metering can help overcome rev-enue recovery challenges, particularly for the types of customers that are least profitable tothe electric utility on a monthly billing model. While these customers benefit from lenientenforcement of bill payment, they also generate an externality on other customers by under-mining the revenue base necessary for infrastructure expansions and maintenance. Prepaidmetering makes substantial progress in narrowing the profitability gap between richer andpoorer customers in our sample, with implications for expanding access in other settings.The heterogeneity we observe also suggests that prepaid metering will be relatively morebeneficial in settings with a more delinquent customer base, a smaller margin per kWh orhigher borrowing costs than Cape Town. Focusing on the City of Cape Town presents a highhurdle for prepaid metering to show positive returns. Unlike in many developing countries,electricity in Cape Town is well managed, with revenue in excess of costs in most years (Cityof Cape Town 2016). Electricity losses are low and problems of bill payment delinquencyand non-payment are comparatively minor. Furthermore, this setting allows us to focus onthe impacts of prepaid metering in an environment free from confounds associated with newconnections or unreliable supply.

Our results highlight the importance of considering the customer response to a change inmetering technology. We speculate that a number of mechanisms may underlie the changes inconsumption that we observe. First, prepaid meters enforce payments automatically and soincrease the experienced price of electricity relative to postpaid metering (even though tariffsremain the same), particularly where enforcement is lax. Second, by forcing customers to payin advance, prepaid metering may affect expenditure patterns. On the one hand, postpaidbilling provides a form of credit and helps smooth income. On the other hand, a high marginalcost of savings together with variable monthly bill totals may lead to debt accumulation bycredit constrained customers. Third, the change from postpaid to prepaid metering transferssome transaction costs from the utility (meter reading, billing) to the customer (purchasing,monitoring consumption). Finally, a number of informational and behavioral differencescharacterize the change from postpaid to prepaid metering. For example, in-home displays

4

provide feedback on usage, more frequent purchases make expenditures more salient, andthe prepaid system offers a form of self- and intra-household control. A clean accountingof these mechanisms is outside the scope of this study. Instead we focus on the impact ofprepaid metering, relative to postpaid billing, on costs and benefits to the utility.

Our findings offer the first evidence on the impact of prepaid metering in a developingcountry context, where they are rapidly becoming the standard technology for new electricityconnections. Existing work on the effects of prepaid electricity metering is scarce, andconsists largely of descriptive studies (Tewari and Shah 2003; Baptista 2013). A recent paperon the largest prepaid metering program in the United States, the Salt River Project, foundreductions in consumption of around 12 percent per month after customers voluntarily switchto prepaid metering (Qiu et al. 2016). The authors rely on a matching design to compareprepaid and postpaid customers, and do not calculate payoffs to the utility. We are notaware of any other plausibly causal evidence on prepaid electricity metering impacts.4

We also contribute to a small but growing body of literature on utility metering andrevenue recovery in developing countries. For example, in a recent paper, McRae (2015)documents the heterogeneous impacts of metering (as opposed to a fixed monthly fee) onhousehold welfare and utility revenue in Colombia. In a study of water bill payment inSouth Africa, Szabó and Ujhelyi (2015) show that an information intervention increasesbill payment rates. More generally, non payment of bills and taxes undermines revenuegeneration in developing countries, and is often associated with challenges monitoring andenforcing tax payments (Gordon and Li 2009; Besley and Persson 2013). A growing number ofempirical studies on taxation show that increasing information for monitoring or changing theincentives associated with enforcement can increase revenue (Kumler et al. 2013; Carrillo etal. 2014; Pomeranz 2015; Khan et al. 2016). Our results echo the conclusions in the taxationliterature that innovations that shift the enforcement burden onto the payee may improverevenue even in settings where detection and enforcement might otherwise be difficult.

The paper proceeds as follows. In the next section, we provide background on prepaidelectricity meters and on the study setting. Section 3 describes the data set and the empiricalstrategy, Section 4 presents the empirical results, and Section 5 calculates the costs andbenefits to the utility. The final section concludes.

4Gans et al. (2013) study the effect of a change in the meter interface that provides prepayment customersin Ireland with additional feedback about real-time usage. The change in feedback occurs for customersalready on prepayment meters and therefore does not provide an independent estimate of the effect ofprepayment on usage.

5

2 Background and context

We begin with a general description of prepaid electricity metering before turning to some ofthe specifics of our setting, including electricity tariffs and billing in the City of Cape Townand details of meter replacement program that we study.

2.1 Prepaid electricity meters

Prepaid electricity meters work on a debit basis: customers purchase electricity and load iton to their meter. As long as the meter has a positive balance, current flows through it intothe home. Once the balance reaches zero, the current is interrupted. Prepaid meters displaythe number of units (kWh) remaining on the meter, and many have features to inform thecustomer when the balance is getting low, such as colored or blinking lights or an audiblealert; they are also generally located within the dwelling in a visible and easy-to-reach place.Contrast this with what we refer to as postpaid metering. On a conventional postpaid meter,customers consume electricity and periodically have their meter read by an employee of theelectric utility. Bills are sent based on cumulative consumption as recorded by the meterreader and customers typically have an additional grace period before their bill is due. Thephysical meters used in most postpaid systems are located outside of the home, and displayconsumption in a way that is difficult for the consumer to access and understand.5

Prepaid electricity metering is attractive to the electric utility for few reasons. First,it generates revenue in advance of consumption and cuts down on non-payment or latepayment of electricity bills. Second, it eliminates the need to send meter readers to physicallyinspect meters, and therefore addresses shirking or bribery, in addition to labor costs andsafety concerns in many settings. Third, it eliminates preparing and mailing monthly billsand processing incoming payments. These last two benefits to the utility are achieved bytransferring some of the transaction costs to the customer. The prepaid system is not withoutcosts to the electric utility, however. There is a substantial initial cost of developing a vendingnetwork to track and charge customers.6 Vendors typically earn a commission on sales ofprepaid electricity, which may be passed on to the consumer or deducted from revenue tothe utility (the latter in our setting). Prepaid electricity meters may also affect theft, which

5Recent innovations with in-home displays and dynamic pricing have begun to change the informationfeedback on some postpaid systems (e.g., Jessoe and Rapson (2014)).

6How sophisticated this needs to be depends on the tariff structure. Under an increasing block tariff,cumulative purchases need to be tracked over the calendar month via a centralized server. We discuss thisin the context of Cape Town’s tariffs below.

6

is unlikely to be an important factor in our setting.7

Customers purchase prepaid electricity from physical or electronic vendors, includingsupermarkets, small shops and kiosks, ATMs, gas stations, and online or via mobile phone.On the vending system used in Cape Town, the customer provides a meter number and themonetary value or number of kWh that they would like to purchase. The vending systemissues an encrypted, meter-specific code based on the kWh purchased that the customerenters into a keypad on the physical meter. The meter itself does not communicate directlywith the grid.8 See Appendix figure A.1 for an example of receipts from Cape Town. Overtime, prepaid metering technology has improved dramatically, both for reliability and theftprevention. The prepaid meters that we study are known as split prepaid meters, becausethe actual meter is located outside the home in a locked kiosk, with only the display andkeypad inside of the home. Communication between the meter and the in-home display is bywire or radio frequency. This design minimizes the risk of tampering and allows the utilityto perform maintenance more easily.

Prepaid metering is expanding across Africa and South Asia. Already widespread inNigeria, Rwanda, Kenya, Uganda, Zambia and elsewhere, and is poised to become the dom-inant metering technology in Africa (Northeast Group 2014). Several countries in Africahave announced that all new connections will use prepaid electricity meters. Prepaid metersare also found in developed countries including New Zealand, the UK and Northern Ireland.A small number of pilot programs have been testing the introduction of prepaid metering inthe United States (see Qiu et al. (2016), for example).

2.2 Electricity in Cape Town, South Africa

In 1990, as South Africa began the transition to democracy, less than a third of SouthAfricans had access to electricity. By the end of that decade the figure had doubled (Bekkeret al. 2008), and by 2011 roughly 80 percent of South African households were electrified(IEA 2013). This extremely rapid expansion of electrification was facilitated in part bythe introduction of prepaid electricity metering, which helped manage revenue recovery for

7Generally, prepaid meters are used to combat theft because innovation in the technology have made itmore difficult to tamper with the meter. However, theft may also increase if customers shift from nonpaymentto illegal connections.

8This description covers the most common type of prepaid metering systems at this point in time. Earliergenerations of prepaid meters were coin-operated or relied on a physical card or key. The technology usedby the City of Cape Town is the STS system, developed in South Africa in the 1980s. It is currently usedfor most prepaid systems around the world.

7

previously unelectrified households (Bekker et al. 2008). In the City of Cape Town, wherewe focus, electrification rates in formal settlements are over 99 percent.

The national state-owned utility, Eskom, owns and operates most generation sources, aswell as the national grid. It sells power in bulk to municipalities, including Cape Town,which pay time of use rates that vary by time of day and month of the year. The City ofCape Town supplies power to roughly 80 percent of residents of the city (the rest are suppliedby Eskom), around 450,000 (75 percent) of whom are on prepaid metering.

The City of Cape Town did not charge a fixed service fee in the years we study, so theincreasing block tariffs are set to cover both fixed and variable costs of electricity supply.9

The tariffs charged on the prepaid meter are the same as on postpaid metering. On aprepaid meter, customers move up the tariff blocks based on cumulative purchases duringthe calendar month; the tariff resets on the first of each month. Over most of the studyyears, two tariffs are used. Residential customers that consume below a threshold quantityof electricity in a 12 month rolling window are on what is referred to as a “Lifeline” tariff,which provides free electricity for up to the first 60 kWh of consumption in a calendar month.Customers not on the Lifeline tariff are charged on a comparatively flat increasing block tariff(“Domestic” tariff). Figure 1 shows the tariffs for 2014-15 in 2014 USD. Tariffs for the otheryears in our data are shown in Appendix Figure A.2. Tariffs are updated each July.

Customers of the City of Cape Town historically received individual bills for each service(water, electricity, refuse removal, etc.) from the City. Over the last decade, customers havebeen shifted to a consolidated billing model, which includes all utilities on a single bill. In oursample, roughly two-thirds of customers received a consolidated bill prior to the switch toprepaid metering. A sample consolidated bill is shown in Appendix Figure A.3. We discussthe implications of consolidated billing for our results in subsequent sections.

2.3 Meter replacement program

In late 2014, the City of Cape Town initiated a program to replace postpaid meters withprepaid meters in selected areas. Suburbs with a low penetration of postpaid meters and a lowaverage property value were targeted, with the idea that eliminating the final few postpaidmeters from these areas would cut out entire meter reading routes. The project consistedof a pilot followed by two stages of implementation. Stage 1 targeted 2,251 postpaid metersin a single suburb called Mitchell’s Plain between November 2014 and February 2015, with

9In other words, the average marginal price exceeds the average marginal cost, and the difference is usedto cover maintenance, new infrastructure and other fixed operating costs.

8

successful replacement of over 90 percent of the targeted meters. Stage 2 targeted 1,995postpaid meters spread across 14 different parts of the City between February and April2015. Replacement rates were lower in the 2nd stage, because compliance became voluntarybeginning in April 2015.

The meter replacement program proceeded as follows. A contractor hired to complete themeter replacements worked in geographically contiguous groups of customers identified bythe City of Cape Town. Based on customer addresses, the contractor first delivered noticesto targeted households informing them that they would have their postpaid meter replacedwith a prepaid meter. Customers were instructed to call to schedule an appointment. Theletter described a time window for scheduling and informed customers that if they had notscheduled an appointment within 15 days, that their electricity would be disconnected. Thiswindow was eventually extended for an additional 15 days. When a customer called, thescheduling window available to them was determined by the order of their geographic groupand by contractor availability. Most meter replacements in a group occurred over a periodof a few days and involved multiple contractor teams. We discuss how this process is usedin our empirical design in Section 3.3.

2.4 Neighborhood characteristics

Our sample consists of 2,251 customers from the suburb of Mitchells Plain and 1,995 cus-tomers from 14 other areas in Cape Town (see the map in Appendix figure A.5). We describethe characteristics of the neighborhoods in our study, based on the 2011 South African cen-sus.10 We begin by describing the Stage 1 sample in Mitchells Plain, a lower to middle incomeneighborhood in the Cape Flats area of Cape Town. During apartheid it was designated asa colored area, and residents were largely excluded from higher paying jobs and received lessaccess to education as a result of apartheid policies.11 Average monthly income is less than300 USD for 42 percent of Mitchells Plain residents, which is close to the average for theCity of Cape Town. The median household spends 8-10 percent of its monthly income onelectricity. Unemployment rates among working age adults are around 32.5 percent, whichis higher than the City average. Electrification is nearly universal: 99 percent of the 38,403households in the 2011 census used electricity for lighting. Nearly all households (92 percent)

10Unless otherwise indicated, all figures are the authors’ own calculations from Statis-tics South Africa’s Census 2011 Community Profile data sets (version 1 from DataFirst -https://www.datafirst.uct.ac.za/dataportal/index.php/catalog/517/get_microdata) and the small arealayer GIS data set for this census (available, upon request, from Statistics South Africa).

11“Colored” is a designation for people of mixed ethnic origin in the South African census.

9

in Mitchells Plain live in formal dwellings and owner occupancy rates are high.12

Customers in Stage 2 of the project are located in areas that are similar to MitchellsPlain on most dimensions. The other areas are, on average, slightly poorer than MitchellsPlain: unemployment was 35 percent, on average, in 2011, and 51 percent of households hadmonthly incomes below 300 USD. Electrification rates and formal property rights are high,like in Mitchells Plain. Overall, rates of electricity use are high in the study sample relativeto low income consumers in other developing countries. Household survey data from projectparticipants in Mitchells Plain indicate that the average household owns a refrigerator, anelectric hot water heater and a television, and cooks and heats using electricity.

3 Data, study design and empirical strategy

3.1 Data

We obtain data from the City of Cape Town’s billing system and prepaid vending systemunder a non-disclosure agreement. Here, we summarize key features of the data. AppendixA.2 contains further details on dataset and variable construction.

3.1.1 Billing data

The City of Cape Town follows a consolidated billing model for most customers (65 percentof our sample at the time of the program), and provides a single bill that covers electricity,water, refuse, sewerage, property taxes and debts (see Appendix Figure A.3 for a samplebill). Bills are sent to customers approximately every 30 days and bills are due 25 days afterthe posting date. Electricity charges are based on physical meter readings taken every 25 to35 days. We use meter reading dates to construct an average daily kWh per month variablethat assumes a constant rate of consumption per day between meter readings. We then takethe average across all days in the month, some of which may have come from different bills.We also construct a measure of the corresponding amount owed for consumption in eachmonth, based on the tariff schedule.

The bill also indicates the date and value of any payments made since the last bill. Weuse this information to construct measures of the days after consumption that a bill is paidoff based on cumulative payments, and an indicator for late payments. We construct threeother measures of delinquency. First, we calculate the share of all monthly bills on postpaid

12Specifically, 88 percent of program participants surveyed in Mitchells Plain report owning their home.

10

metering that were paid past the due date. Second, we construct an indicator for whetherthe customer had multiple outstanding bills at the end of the panel. Third, we constructa measure of whether a customer was ever disconnected while on postpaid metering, andinclude the cost to the City of disconnecting and reconnecting the customer in our benefitcost analysis.

3.1.2 Prepaid vending data

The prepaid vending system records electricity purchases. We use transaction dates toconstruct an average daily kWh variable that assumes a constant rate of consumption perday between transactions. We then take the average across all days in the month, some ofwhich may have come from different transactions. This averaging assumes that customersare not accumulating electricity credit on their meter. As with the postpaid billing data, wealso construct a measure of the corresponding amount owed in month, based on the tariffschedule. We calculate the days between consumption and payment (which will be 0)using the purchase date and the midpoint of the days before the next purchase.

3.1.3 Marginal costs and revenue recovery costs

We obtain data on the marginal cost of electricity supply from the City of Cape Town.The City purchases electricity from Eskom, the national provider, at tariffs that vary bytime of day and month of the year. We use the city-wide monthly average price per kWhpaid to Eskom as our marginal cost of supply measure (see Appendix figure A.4). TheElectricity Department has an estimated loss rate of 11.25 percent, which includes bothtechnical and non-technical losses, which we assume both meter types equally. Prepaidelectricity vendors receive approximately 0.002 USD per kWh sold, which is deducted fromthe revenue remitted to the City. We also gather information from the Electricity Departmenton other administrative costs associated with revenue recovery on postpaid billing, includingthe average per-customer meter reading and bill preparation costs.

3.2 Electricity outcomes across metering types

The billing data and prepaid vending data measure electricity in different ways. As describedabove, the billing data are used to construct a daily average between bill t and t�1, while theprepaid dataset is used to construct a comparable measure between transaction t and t+ 1.

11

The constructed average daily consumption per month removes differences in the frequencyof observations across the dataset.

Given that the prepaid metering system records transactions, rather than actual con-sumption, the comparability of the measures across systems deserves further discussion. Akey assumption in transforming expenditures into a measure of average daily consumptionis that customers are not accumulating unused electricity on their meter. Three observa-tions suggest that this is reasonable. First, the increasing block tariff structure provides anincentive not to accumulate unused electricity since tariffs reset on the first of the monthas discussed above. Second, customers tend to purchase very small amounts of electricity,very frequently: the median purchase size is 19 kWh or around 30 ZAR (2.6 USD), andthe median purchase frequency is every 3 days. Finally, while the prepaid technology mightlead to greater smoothing if customers have a tight monthly budget constraint, we note thatthe within-customer coefficient of variation of our preferred constructed average daily kWhmeasure is very similar for prepaid and postpaid observations (0.24 for postpaid and 0.21for prepaid) and even more similar when the outcome is in logs (0.11 versus 0.10). Theassumption that customers are not accumulating unused electricity on their meter may bemost difficult to satisfy in the month or two following the meter replacement as customersbuild up a minimum balance on their meter. Our main analysis drops the month of meterreplacement. We show an event-study style analysis that shows how electricity use evolvesin the months following the meter replacement.

More generally, the transformation of expenditure data into a measure of consumptionis a common challenge in studies of household demand. In the case of prepaid electricity, weconstruct a monthly measure of consumption based on a mean (median) purchase frequencyof 12.6 (9.5) times per month. Over 95 percent of customers make at least one purchaseevery month on prepaid metering. While a number of econometric methods have beenadvanced for estimating latent demand based on less frequent purchases (e.g., Kay et al.(1984); Blundell and Meghir (1987)), the high frequency of observed expenditures togetherwith the administrative nature of our data allows us to make the simplifying assumptionthat monthly expenditures and monthly consumption converge.

3.3 Study design

We collaborated with the City of Cape Town to randomize the phase in of the meter instal-lation program. Groups of around 150 adjacent customers were randomized in two separate

12

waves, one for each of the meter replacement stages. First, in Mitchell’s Plain (stage 1),we imposed a grid over a map of all targeted customers, and aggregated cells until around150 adjacent customers formed a randomization group. Thirteen groups were identified andtheir order was randomized. Second, targeted customers in the other suburbs (stage 2) wereaggregated in a similar process, but used pre-existing suburb definitions to assign customersto groups. The second stage included 14 groups, which were also randomly ordered. Thus,the final design consists of 27 groups, with separate random assignment within the first andsecond groups. A map of the randomization groups is shown in Appendix figure A.5.

3.3.1 Potential confounds

Two issues associated with implementation present potential confounds. We discuss these inturn, anticipating the robustness checks that we perform in Section 4.1.2. First, when theCity installs a new meter, it is defaulted into the Domestic tariff. If the meter should beassigned to Lifeline tariff, the change has to be manually entered into the billing or vendingsystem. Because of the volume of meter replacements around the time of the project, therewere some delays in restoring the Lifeline tariff to customers who had received it prior to theswitch. Specifically, 981 of the 1,335 customers on Lifeline in the month before the switchwere put on to the Domestic tariff for at least a portion of the month following the switch.The majority of these cases were corrected within three months of the switch, however,some lasted considerably longer, and some may indicate a permanent tariff change for thecustomer. A tariff change that coincides with the switch to prepaid metering presents anobvious confound to identifying the effects of the meter on electricity use. We control forerroneous tariffs in all of our analyses, and drop these customers altogether in a robustnesscheck. Our analysis of heterogeneous treatment effects show separate results for Domestictariff customers, who were unaffected by the tariff error.

Second, around the same time as the meter replacement program, the City of CapeTown implemented several months of rolling blackouts (loadshedding) to manage supplyshortages. These involuntary outages may lead to a reduction in electricity use that isunrelated to prepaid metering. We obtain data on the hours of loadshedding and otheroutages per month for each substation serving Mitchell’s Plain, where meter installationswere most likely to overlap with the period of relatively intense loadshedding. Beginningin November 2014, substations in Mitchell’s Plain experience an average of 1 to 1.5 hoursof load shedding per month, or around 0.04 percent fewer hours of electricity supply (seeAppendix figure A.6). We construct controls for the average hours per month of outage

13

across all Mitchell’s Plain substations, and control for them in a robustness check.13 We alsoperform robustness checks that include comparison households outside of the project samplewho were subject to loadshedding but not meter replacements.

3.3.2 Study sample

The final sample for analysis is based on the lists of targeted customers used in the ran-domization. Table 1 shows characteristics of the sample. We observe 4,246 customers, withan median of 54 monthly observations per customer, 16 of which are from a prepaid meter.On average, customers use around 16 kWh of electricity per month prior to the program,and owe around 52 USD per month for electricity. Over half of their bills are paid late, andit takes an average of around 100 days to pay after the date of consumption. Twenty-sixpercent of customers have multiple outstanding unpaid bills at the end of the panel and 21percent were ever disconnected as a means of enforcing bill payment. The median propertyvalue is around 27,000 USD and 31 percent of customers are on the subsidized tariff at thetime of the program.

3.4 Empirical strategy

We observe electricity and other billing outcomes for a customer i in each month-year t. Toidentify the effect of a switch from postpaid metering to prepaid metering on outcome y

it

,we estimate:

y

it

= ↵ + �prepaid

it

+ ⌧

t

+ tari↵it + ⌘

i

+ ✏

it

(1)

where prepaid

it

indicates the share of month t that customer i received electricity througha prepaid meter, ⌧

t

are time fixed effects (either month-year or separate calendar monthand billing year), tari↵it is a time varying Lifeline tariff indicator, and ⌘

i

are customerfixed effects.14 Our main results exclude the switch month, resulting in a binary measureof prepaid. Standard errors are clustered at the customer level. As a robustness check, wecompute t-statistics based on standard errors clustered at the randomization group level.Under the assumption that prepaid metering is uncorrelated with ✏

it

, conditional on time13We do not obtain the data for the suburbs in stage 2 of the meter replacement program. The loadshedding

schedule was similar across suburbs and adhered to fairly closely.14We also add an indicator for months affected by the errors updating the Lifeline tariff in the months

following the meter switch.

14

and customer fixed effects, � identifies the causal impact of prepaid metering on outcomey

it

. We also estimate the effect of prepaid metering over time through an event study styleanalysis that recovers a separate treatment effect for each month pre- and post- receivingthe prepaid meter.

While meter replacement was involuntary, the actual meter replacement date may notbe not exogenous to electricity consumption. First, there is a non-trivial amount of non-compliance, particularly in stage 2 of the project. Second, the timing of replacement evenwithin the assigned windows may be determined by factors correlated with use, such ascustomer work schedules or resistance to the replacement. In exceptional circumstances,customers were allowed to make appointments outside of the assigned replacement window,with replacements carried out by a single contractor team. Third, in practice, the length oftime that the contractor spent in each group was determined by the difficulty in schedulingall appointments. Our primary instrument for assignment to a prepaid meter is thereforethe first date on which the contractor has more than one team working in the randomizationgroup, which corresponds to the start of the scheduling window. As an alternative, we alsouse the date on which the first households in the group received a notification (maildrop)about the project.

Figure 2 summarizes the two instruments, and the actual switching patterns. The ran-domization groups are ordered on the vertical axis, with the Mitchell’s Plain groups (stage1) corresponding to groups 1-13. The circles indicate the timing of the first letters deliveredin the group. The squares show the first date on which the contractor was performing in-stallations in the group, and the triangles indicate the median switch date among those whoever switched, with the share ever switched printed on the figure. As the figure illustrates,letter deliveries followed the random ordering of the groups very closely, particularly duringstage 1 of the program. The contractor dates are also consistent with the randomized order,though they occasionally deviate, particularly toward the end of the program. We use thecontractor date as our main instrument since it most accurately reflects the exogenous deter-minant of take up. We show results using the group order and mailing dates as instrumentsin robustness checks. Given the higher compliance rates overall and the better adherenceto the randomization, we report our main results for Mitchells Plain alone in a robustnesscheck. The IV estimates estimate the local average treatment effect of prepaid metering byusing the assignment variables as instruments for receiving a prepaid meter.

Recovering the causal effect of prepaid metering, either through the OLS or IV estimates,requires that the actual order or assigned (randomized) order of meter replacements is un-

15

correlated with unobserved time-varying factors that affect electricity consumption. Timeinvariant customer characteristics are absorbed in the customer fixed effects used in all anal-yses. Appendix table A.1 shows the correlation between pre-switch customer characteristics(prior to November 2014) and the assigned group order (column 2), the assigned switchdate, based on the contractor instrument (column 3), the actual switch date (column 4)and whether the customer was ever switched (column 5). The small number of groups in-cluded in the randomization results in a lack of balance on property value (and thereforeother characteristics, presumably), which would be of greater concern if the data did notallow for household and time fixed effects. We examine pre-program residuals for “paralleltrends” after removing month-year and household fixed effects from the relationship betweenaverage daily kWh and month, prior to the program. We plot the average residual withineach randomization group for the 34 months leading up to the start of the program and ob-serve relatively flat trends that do not differ substantially across groups (see Appendix figureA.7). Appendix figure A.8 also plots monthly average daily kWh for the project sample anda comparison group of postpaid customers elsewhere in the City, and shows parallel trendsin the months preceding the project start.

4 Results

We start by showing the impact of switching to a prepaid meter on electricity use. Second,we turn to heterogeneity in the treatment effects by customer characteristics and behaviorprior to the replacement program. Finally, we estimate impacts on variables that inform thebenefit cost analysis, including amount owed and payment patterns.

4.1 Electricity use

Figure 3 plots the median daily average consumption in each month and the share of tar-geted customers switched to prepaid metering. The figure clearly shows the strong seasonalpattern of consumption, which peaks in the South African winter, when many customers useelectricity to heat their homes. The drop in usage that coincides with the prepaid meteringprogram is clearly visible in the figure.

Table 2 shows the regression coefficients from OLS regressions (columns 1 and 2) and2SLS regressions using the contractor date instrument (columns 3 and 4). The outcomein Panel A is average kWh/day and in Panel B is the log of average kWh/day. Columns

16

1 and 3 show the results with separate month and year fixed effects and columns 2 and 4use month-year fixed effects. The OLS coefficients (columns 1 and 2) indicate that averagedaily usage fell between 1.89 and 2.17 kWh (Panel A), depending on the specification. Thiscorresponds to a decline of between 11.9 and 15.2 percent (Panel B). The IV and OLS resultsare generally similar in magnitude, with slightly smaller results from the IV in levels (PanelA). . The first stage on the IV regressions is – unsurprisingly – very strong (F-statistic >1000). The reduction is precisely estimated in all specifications.15

4.1.1 Persistence

We next provide suggestive evidence of impacts over time. We set the month prior to theswitch as time 0 and estimate separate coefficients for each month following the switch (theswitch month is omitted).16 Figure 4 plots the resulting coefficients. Overall, the reductionpersists for the first year on prepaid metering. If anything, the point estimates appear todecline over time. Note that customers that do not switch meter types are not included inthe event study results.

4.1.2 Robustness checks

To examine the robustness of our estimate of the effect of prepaid electricity metering onelectricity use, we repeat the analysis above with alternative specifications, including differentinstruments and a difference in difference design that includes comparison customers. Asshown in Appendix table A.4, the coefficient remains reasonably stable, implying that theresults are not driven by outliers or by the assumptions used in variable construction.17

Appendix figure A.8 shows the descriptive consumption plot underlying the main differencein difference specification (column 4).18 Next, we examine the robustness to alternative

15We also construct t-statistics based on standard errors clustered at the randomization group level. Forour main specifications, we perform 1000 iterations of the wild cluster bootstrap procedure (see Cameronet al. 2008), none of which generate t-statistics below the value without clustering, implying a p-value of<0.0001.

16We include month and billing year fixed effects, and cluster standard errors at the customer level. Notethat we cannot simultaneously control for time and household fixed effects and identify the event studycoefficients, so we omit household fixed effects from the specification.

17Differences across the coefficient on the alternative instruments is unsurprising given that the marginalcustomer is likely to differ depending on which instrument is used. Nevertheless, the results are qualitativelysimilar across all instruments.

18Note that neither of the comparison samples are ideal. The sample of postpaid comparison customers(column 4) comes from other parts of the city since all postpaid customers in the project neighborhoodswere targeted for meter replacement. The sample of prepaid comparison customers is drawn from the

17

sample restrictions and manipulations of the outcome variable, and again see little differencein the point estimates (see Appendix table A.5). Finally, we implement two robustness checksdesigned to test remaining concerns about the implementation issues discussed in Section3.3.1; we control for loadshedding and exclude Lifeline customers that received the incorrecttariff when their meter was replaced. Results change very little (see Appendix table A.6).Finally, we conduct a placebo test that reassigns the switch date and the instrument to oneyear prior to its actual occurrence and observe a small and insignificant coefficient on theprepaid indicator (see Appendix table A.6, column 4).

4.2 Heterogeneous treatment effects

We preview the heterogeneity that we will return to in the benefit cost analysis (Section 5),by examining impacts by observable customer characteristics. We re-estimate equation (1),interacting the prepaid indicator with month and billing year fixed effects with each binaryheterogeneity measure, summarized in Table 3.19 Columns 1 and 2 show the OLS effecton average daily kWh and columns 3 and 4 show the results instrumenting for the switchdate with the first date the contractor worked in the group. We show the total effect foreach subgroup in columns 1 and 3 and the difference (i.e. the coefficient on the interactionterm) in columns 2 and 4. We also report the number of customers and the postpaid averagedaily kWh for each sub-group. We show the results in levels and report pre-switch averageconsumption by subgroup in the column labeled Mean. The pairwise correlations betweenthe heterogeneity variables is shown in Appendix Table A.3.

We begin with customer tariffs at the time of the program. On average, Lifeline customersuse less electricity per month than Domestic customers, and pay less per kWh (see Figure 1).The first two rows Table 3 show results by tariff prior to the switch. Domestic customers cutback significantly more than do Lifeline customers. Proportional to their average consump-tion on postpaid, the reduction by domestic customers is also larger, at around 12 percentrelative to a reduction of around 10 percent for Lifeline customers. Domestic customersalso tend to consume more electricity than do Lifeline customers; unsurprisingly, when wesplit the sample into above and below median postpaid consumption, we see similar results,with larger absolute (and proportional) reductions among the larger consumers. To examine

same neighborhood but provides a different counterfactual: what consumption would have been if projectcustomers had always been on prepaid metering.

19We focus on the specification that includes separate month and year fixed effects because the prepaidcoefficient may not be well identified within month-year for all sub-groups.

18

whether these two sets of results are simply capturing differences by wealth levels – Lifelinecustomers and below median consumers also tend to be poorer (Appendix Table A.3) – wesplit the sample by property value, which is the wealth proxy that we observe in the data.We divide customers according to the valuation used for many other anti-poverty programsin the City of Cape Town (300,000 ZAR based on the 2012 valuation, which corresponds toaround 29,000 USD2014). We see larger reductions among low property value customers,both in levels and proportional to their average consumption on postpaid metering.

Next, we examine responses by three measures of delinquency. First, customers that paymore than the median number of bills late (more than 58 percent) prior to the programreduce their consumption by more than do customers who tend to pay their bills on time.Next, customers with multiple outstanding unpaid bills at the end of the panel, and thosethat experienced disconnection at some point as a means of payment enforcement bothshow similar reductions to customers with a better history of payment (if anything, theyare slightly less responsive). These last results are particularly relevant for considering thepayoffs to the utility from switching customers from postpaid to prepaid billing, which weturn to in the next section.

4.3 Payment behavior and cost of supply

We turn from these impacts on kWh/day to the measures that will feed into the utility’sbenefit cost analysis, specifically, the amount owed by the customer, the cost of supply andthe timing of customer payments. The outcomes we show here exclude administrative costssuch as bill preparation, which we include in the benefit cost analysis and explain in thenext section. Summary statistics for each of these outcomes prior to the meter replacementprogram are shown in Figure 1.

Effect on amount owed per month Electricity is priced on an increasing block tariff,which means that reductions in consumption may lead to proportionally larger reductionsin the amount owed from customers whose reduction crosses a tariff step, or proportionallysmaller (zero) reductions for customers that only consume their free electricity allowance oneither meter type. We estimate the impact of prepaid metering on the amount owed forconsumption in a month (Table 4, column 1).20 Customers owe 6.8 to 6.9 USD less per

20This amount owed variable is calculated based on a month-level version of our main average kWhmeasure, multiplied by the customer’s marginal price on each tariff block.

19

month on prepaid metering. This corresponds to a reduction of around 13 percent eachmonth relative to the average amount owed on postpaid metering (see Table 1).

Effect on marginal supply cost The utility purchases electricity from the national elec-tricity company based on prices that vary with time of day and month of the year. Whilewe do not observe the time of the day of consumption, we do observe the month of the year.Reductions in consumption are most valuable if they occur in June, July and August, whenthe utility pays a higher marginal cost per kWh (see Appendix figure A.4). We estimateseparate treatment effects on the average kWh/day for each calendar month, and plot theresults in Figure 5. These indicate that the largest reductions occur in the months whenper kWh costs to the utility are highest. Column 2 of Table 4 shows the impact on themarginal cost of supply, calculated by multiplying the utility’s marginal cost in the monthby the customer’s total kWh in the month. The point estimates indicate a 2.8 to 2.9 USDper customer per month reduction (column 2), which is considerably below the reduction inamount owed reported in column 1.

Payment timing The customer response to prepaid metering consists both of the quantityof electricity consumption and when it is paid for. Mechanically, prepaid metering moves thepayment date from after consumption to before consumption. In addition, the billing cycleadds a delay of around 60 days to the time between when a postpaid metering customerconsumes the average kWh in a month and when payment is due for that kWh. As discussedin Section 3.1, our measure of payment timing is calculated as the days between whenconsumption occurs and when payment arrives to the utility. Column 3 of Table 4 shows theimpact on the days between consumption and payment. Prepaid metering results paymentthat arrives 75 (OLS) to 85 (IV) days sooner. Note that this measure of payment timingexcludes payments that never arrive; i.e. average payment timing excludes bills that areeventually written off.

5 Costs and benefits to the utility

We calculate the costs and benefits of the meter replacement program for the City of CapeTown, based on our empirical estimates and administrative cost records. We then examineheterogeneity and implications for other settings.

20

5.1 Present value of the meter replacement program

We write down a simplified present value expression for each metering type m = {pre, post}to summarize the costs and benefits to the utility in each month t:21

PV

m =pq

m

(1 + i)t+s

m

� cq

m + b

m

(1 + i)t. (2)

Calculation of the present value requires estimation of three parameters: (1) the revenuepaid to the utility (pqm), (2) the timing of payment (sm) and (3) the cost of supply andrevenue recovery (cqm + b

m). We discuss each in turn.

Revenue The customer consumes quantity q

m in month t. We account for the increasingblock tariff structure by calculating the total amount owed (pqm) for consumption in montht. In our setting, tariffs are the same across metering technologies, though the utility sharesa small fixed margin per kWh (0.002 USD) with vendors on the prepaid system, regardlessof the price paid by the customer. This revenue measure is the same as what we calculate asthe amount owed for consumption in Section 4.3, except that it is net of vendor margins andis set to zero in the case of non-payment, which is is defined as debts that are not recoveredwithin three years of the month of billing.22

Timing of payment The value of consumption to the electric utility also depends on whenthe customer pays, sm, which is expressed in months relative to the consumption month t.On prepaid metering, spre 0 since the customer cannot consume until they have purchasedelectricity. On postpaid metering, spost > 0 since the customer does not receive a bill untilsometime after consumption occurs. This measure is the same as what we show in Section4.3, expressed in months instead of days.

21These calculations avoid a more complete modeling of consumer and utility decisions. For example,utility decisions about enforcing non-payment affect both revenue recovery rates and revenue recovery costs.On the customer side, we take point estimates from our impact evaluation as given and do not model themechanisms that relate the metering technology to consumption. In addition, we ignore capital and fixedsupply costs, including the meter itself. A prepaid meter costs around USD 66, while a postpaid meter costsaround USD 30.

22Debts older than three years are officially written off by the City. To calculate payment probabilities forunpaid bills at the end of our panel that are younger than three years, we calculate a “hazard rate” or theprobability of default given that no payment has occurred by the number of months since billing. Overall,we estimate that around 1.9 percent of the revenue owed on the postpaid system is never recovered, whichis consistent with the estimates used by the City of Cape Town, but is much better than revenue recoveryrates for most developing country utilities. See Section 3.1 and Appendix A.2 for further discussion of thepayment variables.

21

Cost of supply and revenue recovery The utility’s marginal cost of supply varies bytime of the day and and month of the year, but we observe only average consumption permonth which we use to calculate a monthly kWh supply cost, which is the same as inSection 4.3. We add to this technical and non-technical losses of 11.25 percent, based onestimates provided by the Electricity Department (i.e. the cost of each kWh supplied is 1.1125times the marginal cost). Recurring monthly costs b

m include meter reading costs and billpreparation, which are the same for all postpaid customers in the data, and total 1.74 USDper customer per month. We also observe that around 20 percent of the customers in ourdata are disconnected (and eventually reconnected) in at least one month while on postpaidmetering. Conditional on ever being disconnected, the average number of disconnections istwo. Disconnections are costly to the City, which charges customers only 30 USD of the 120USD it takes to disconnect and reconnect a customer. We include the cost of disconnectionin the month in which a disconnection is observed. Finally, our calculations ignore fixed costsassociated with infrastructure and operation, which are covered out of the rate structure inCape Town. Thus, results should not be interpreted as profits.

Relative returns to the utility The electric utility’s object of interest is the presentvalue of prepaid metering relative to postpaid metering, i.e. PV

pre

/PV

post. We can calculatethis for the City of Cape Town by plugging in our estimates from the meter replacementproject together with administrative costs obtained from the City. A summary of the valuesused is provided in Table 5. The predicted means of amount owed, supply cost and timingof payment variables are estimated for the sub-sample of compliers (i.e. switched customers)using OLS with separate calendar month and year fixed effects, and controls for the tariffand the months with tariff errors (see Section 3.3.1).23 The final rows show the average per-customer monthly present value calculation for the City of Cape Town for an annual interestrate of 8 percent.24 Bootstrapped standard errors are in parentheses. The results showa return that is around 10 percent higher on prepaid metering than on monthly postpaidbilling. Note that this proportional gain ignores the one-time cost of the meter and so is

23For the purpose of analyzing average costs and benefits associated with meter type, we restrict the sampleto compliers only and focus on the policy relevant average treatment effects from the OLS specification asopposed to the local average treatment effects produced by the IV. For most outcomes, the OLS and IVspecifications produce similar results. In addition, Appendix table A.5 shows similar effects on electricityuse for compliers only as for the whole sample.

24The City of Cape Town uses a discount rate approved by the National Energy Regulator of South Africa(NERSA) for project evaluation. From 2010 to 2016, the rate was was 8% and a payback period of 25 years,as stipulated by the South African electricity grid code.

22

constant over time. It takes a little over seven years for the additional revenue to cover thefixed costs of the prepaid meter (USD 66).

We perform back of the envelope calculations to calibrate the share of the gains thatcome from different sources. The value of receiving revenue sooner results in returns toprepaid metering that is around 4.5 percent higher than on postpaid, or nearly half of theoverall increase in relative returns.25 The utility also avoids an expected loss of 0.84 USDper customer per month associated with billing defaults, which amounts to 1.6 percent ofthe revenue owed. On prepaid metering, the utility also avoids billing and meter readingcosts that result in fixed per customer costs per month. On the prepaid system, these arereplaced, in part, by vendor commissions that are proportional to consumption. For monthlyconsumption less than 880 kWh per month (93 percent of customer-months on postpaid),the costs of prepaid vending commissions are less than the costs of meter reading and billpreparation. Finally, the utility avoids an expected enforcement cost of disconnection of0.89 USD per customer per month. The relative return to prepaid metering is considerablybelow what would be expected if the reduction in consumption were ignored. If consumptionis assumed to remain flat, then the difference in administrative costs, together with earlierpayments leads to a relative return that is around 33 percent higher for prepaid metering,i.e. over three times the returns we calculate.

5.2 Heterogeneity and generalizability

The costs and benefits of prepaid metering depends both on the consumption and paymentbehavior of the customer base and also on features of the tariff and cost environment. As aresult the numbers from Cape Town need not generalize to other settings. We extend ouraverage cost benefit calculation to examine how the results vary by customer type and theadministrative setting.

First, we revisit the customer characteristics analyzed in Section 4.2. We calculate esti-mates of the returns from prepaid metering relative to postpaid metering for each sub-group.Note that a relative return of 100 percent implies that prepaid and postpaid meters generatethe same returns to the utility. Figure 6 summarizes the results along with bootstrapped95 percent confidence intervals (see also Appendix table A.7). We include the monthly re-turns to postpaid metering for each customer type on the figure, and observe considerable

25This return is based on equation (2) using revenue and costs for prepaid metering, and sm from eachmeter type based on estimates in Table 5. If instead we calculate the effect of earlier payments using revenueand costs for postpaid metering, the relative returns are around 5.4 percent.

23

heterogeneity by characteristic in the baseline returns to the electric utility for serving eachtype of customer. Each of the characteristics represents a potential targeting rule for theCity of Cape Town, and are suggestive of how results might look in other settings, with adifferent customer base. Relative returns are positive for both tariff types. Low consumersgenerate considerably higher payoffs than do high consumers, whose relative returns are veryclose to zero. Low property value customers, defined by the 300,000 Rand value cutoff usedfor other policies by the City of Cape Town, also generate higher relative returns than cus-tomers with higher property values. Perhaps unsurprisingly, relative returns are positive forthe delinquent customers in our sample: those with late payments, outstanding debts or ahistory of disconnection on postpaid. Less delinquent customers, particularly those with nooutstanding payments at the end of the panel, actually generate negative returns on prepaidrelative to postpaid metering. Notably, the customer characteristics that are associated withthe highest returns from prepaid metering are likely to be considerably more prevalent inother developing country settings, and are the mostly likely types of customers to be rationedon a postpaid billing system.

Next, we look to alternative assumptions about the cost environment. In Cape Town,the discount rate used for project evaluation is relatively low at 8 percent. This places littleweight on the value of revenue received earlier under prepaid metering. The top panel ofFigure 7 shows that returns to each metering type fall as interest rates rise, with a steeperslope for postpaid meters. The bottom panel shows returns by multiples of the marginal costof supply. Losses on the City of Cape Town’s network are around 11.25 percent, includingboth technical and non-technical losses. Other developing country utilities tend to havehigher loss rates (Trimble et al. 2016), and so may face a higher cost of supply. Alternatively,in settings where costs are high relative to tariffs, the estimates showing higher marginal costssuggest that the relative benefit of prepaid metering will be higher. This is due largely tothe fact that reductions in consumption are less costly to the utility as the marginal kWhbecomes less profitable.

6 Conclusion

We study the impact of prepaid electricity metering on electricity use and utility returnsin Cape Town, South Africa. Using a unique dataset that tracks customers as they areinvoluntarily switched from postpaid to prepaid metering, we document a 13 percent decreasein electricity use, which persists for the 12 months following the switch. The average effects

24

mask considerable heterogeneity, with larger proportional reductions from poorer customersand those with a history of paying their monthly bills late.

Customer responses, both in terms of overall consumption and in the effect of the me-tering change on payment patterns, have implications for utility revenue and cost flows. Wecalculate the returns to the utility from prepaid metering, relative to postpaid metering.On the one hand, the utility recovers the revenue it is owed sooner and more completely.On the other hand, the amount owed decreases substantially as consumption falls. Ourcost benefit analysis for Cape Town also accounts for billing costs, prepaid vendor margins,and technical and non-technical losses on the system. We show that the returns to prepaidmetering will depend on these features of the cost environment, as well as on the customerbase. In our sample, smaller and poorer consumers and more delinquent customers yield thehighest returns to prepaid metering relative to monthly billing. These characteristics arelikely to be shared by customers in other developing country settings, where the decision toconnect a customer may depend on expected returns. Therefore prepaid metering may bean important tool for expanding energy access via the grid to poor consumers.

Our data do not allow for a clear accounting of the mechanisms underlying the customer’sresponse to prepaid metering, though we clearly show that customer responses must be takeninto consideration when assessing the payoffs associated with prepaid metering. Ignoringimpacts on consumption leads to a substantial over-estimate of the benefits to the utility.A full welfare accounting must, of course, consider the impacts on the customer. In manysettings, the choice between prepaid and postpaid metering is likely to affect the extensivemargin of access. Consequently, the proper welfare comparison may be electrification withprepaid metering or no electricity access at all.

25

References

Baptista, Idalina, “Everyday practices of prepaid electricity in Maputo, Mozambique,”Working Paper, 2013.

Bekker, Bernard, Anton Eberhard, Trevor Gaunt, and Andrew Marquard, “SouthAfrica’s rapid electrification programme: Policy, institutional, planning, financing andtechnical innovations,” Energy Policy, 2008, 36 (8), 3125–3137.

Besley, Timothy and Torsten Persson, “Taxation and Development,” Handbook of PublicEconomics, 2013, 5, 51.

Blundell, Richard and Costas Meghir, “Bivariate alternatives to the Tobit model,”Journal of Econometrics, 1987, 34 (1-2), 179–200.

Cameron, A Colin, Jonah B Gelbach, and Douglas L Miller, “Bootstrap-basedimprovements for inference with clustered errors,” The Review of Economics and Statistics,2008, 90 (3), 414–427.

Carrillo, Paul, Dina Pomeranz, and Monica Singhal, “Dodging the taxman: Firmmisreporting and limits to tax enforcement,” Unpublished, 2014.

City of Cape Town, “Municipal Budget, 2015/16 - 2017/18,” 2016.Dinkelman, Taryn, “The effects of rural electrification on employment: New evidence from

South Africa,” American Economic Review, 2011, 101 (7), 3078–3108.Gans, Will, Anna Alberini, and Alberto Longo, “Smart meter devices and the effect

of feedback on residential electricity consumption: Evidence from a natural experiment inNorthern Ireland,” Energy Economics, 2013, 36, 729–743.

Gaunt, Trevor, “Meeting electrification’s social objectives in South Africa, and implicationsfor developing countries,” Energy Policy, 2005, 33 (10), 1309–1317.

Golumbeanu, Raluca and Douglas F Barnes, “Connection charges and electricity accessin sub-Saharan Africa,” World Bank Policy Research Working Paper, 2013, (6511).

Gordon, Roger and Wei Li, “Tax structures in developing countries: Many puzzles anda possible explanation,” Journal of public Economics, 2009, 93 (7), 855–866.

Henderson, Vernon, Adam Storeygard, and David Weil, “A Bright Idea for MeasuringEconomic Growth,” American Economic Review, 2011, 101 (3), 194–199.

Heymans, Chris, Kathy Eales, and Richard Franceys, “The Limits and Possibilitiesof Prepaid Water in Urban Africa: Lessons from the Field,” Technical Report, World Bank2014.

IEA, “World Energy Outlook Electricity Access Database,” Technical Report, International

26

Energy Agency 2013.Jessoe, Katrina and David Rapson, “Knowledge is (Less) Power: Experimental Evidence

from Residential Energy Use,” American Economic Review, 2014, 104 (4), 1417–1438.Kay, John A, Michael J Keen, and C Nicholas Morris, “Estimating consumption

from expenditure data,” Journal of Public Economics, 1984, 23 (1), 169–181.Khan, Adnan Q, Asim I Khwaja, and Benjamin A Olken, “Tax farming redux: Exper-

imental evidence on performance pay for tax collectors,” Quarterly Journal of Economics,2016, forthcoming.

Kumler, Todd, Eric Verhoogen, and Judith A Frías, “Enlisting employees in improv-ing payroll-tax compliance: Evidence from Mexico,” Working Paper, 2013.

Lee, Kenneth, Edward Miguel, and Catherine Wolfram, “Experimental Evidence onthe Demand for and Costs of Rural Electrification,” Technical Report, National Bureauof Economic Research 2016., Eric Brewer, Carson Christiano, Francis Meyo, Edward Miguel, MatthewPodolsky, Javier Rosa, and Catherine Wolfram, “Electrification for "Under Grid"households in Rural Kenya,” Development Engineering, 2016, 1, 26–35.

McRae, Shaun, “Efficiency and Equity Effects of Electricity Metering: Evidence fromColombia,” Working Paper, 2015.

Northeast Group, “Sub-Saharan Africa Electricity Metering: Market Forecast (2014 to2024),” Technical Report, Northeast Group, LLC 2014.

Pomeranz, Dina, “No Taxation without Information: Deterrence and Self-Enforcement inthe Value Added Tax,” American Economic Review, 2015, 105 (8), 2539–2569.

Qiu, Yueming, Bo Xing, and Yi David Wang, “Prepaid electricity plan and electricityconsumption behavior,” Contemporary Economic Policy, 2016, forthcoming.

Szabó, Andrea and Gergely Ujhelyi, “Reducing nonpayment for public utilities: Ex-perimental evidence from South Africa,” Journal of Development Economics, 2015, 117,20–31.

Tewari, Devi Datt and Tushaar Shah, “An assessment of South African prepaid elec-tricity experiment, lessons learned, and their policy implications for developing countries,”Energy Policy, 2003, 31 (9), 911–927.

Trimble, Christopher Philip, Masami Kojima, Ines Perez Arroyo, and Farah Mo-hammadzadeh, “Financial viability of electricity sectors in Sub-Saharan Africa: quasi-fiscal deficits and hidden costs,” World Bank Policy Research Working Paper, 2016, (7788).

van Heusden, Peter, “Discipline and the "new logic of delivery": Prepaid electricity in

27

South Africa and beyond,” Electric Capitalism: Recolonising Africa on the Power Grid,2012, p. 229.

Wolfram, Catherine, Orie Shelef, and Paul Gertler, “How Will Energy Demand De-velop in the Developing World?,” Journal of Economic Perspectives, 2012, 26, 119–138.

28

0.05

.1.15

.2USD/kWh

0 25 60 350 600kWh

Lifeline  1 Lifeline  2Domestic

Jul  2014  -­  Jun  2015  (in  2014  USD)

Figure 1: City of Cape Town electricity tariffs 2014-2015

Notes: Tariff schedules for July 2014 to June 2015. Tariff assignments are determined by a 12 monthrolling average of past electricity use. See text for additional details.

29

0.91

0.88 0.94

0.98 0.95

0.92

0.87

0.86 0.84

0.92 0.73

0.91

0.87

0.55 0.60

0.74 0.85

0.63

0.30

0.60

0.72 0.50

0.73 0.58

0.55

0.50 0.88

14

710

1316

1922

2528

Ran

dom

izat

ion

grou

p

15 Dec 2014 15 Feb 2015 15 Apr 2015

MaildropContractor begins workMedian switch date (share switched)

Holiday

Figure 2: Switching patterns by randomization group

Notes: Project dates by randomization group (groups 1-13 correspond to the Mitchell’s Plain sam-ple). The contractor instrument turns on on the first date that more than one contractor teamis working in the group. The maildrop instrument turns on on the first date that customers inthe group received information about the meter replacement program. The median switch date isconditional on switching and the share of customers that ever switched in the group is printed onthe figure.

30

0.2

.4.6

.81

Shar

e sw

itche

d

1012

1416

18M

edia

n da

ily a

vg k

Wh

2012m1 2013m1 2014m1 2015m1 2016m1

Median daily avg kWh Share switched

Figure 3: Daily average kWh/month and share switched over time

Notes: Median average kWh in each month and the share of targeted customers switched to prepaidmetering. Note that the peak months of June and July correspond to the winter months in CapeTown.

31

-4-2

02

Avg

daily

kW

h (e

st'd

coe

ff)

-12 -6 0 6 12Months since switch

Figure 4: Event study (average daily kWh)

Notes: Event study of the effect of prepaid metering on average daily kWh. The OLS specificationsets the month prior to the meter switch as the base month and regresses average daily kWh onmonths since the switch, conditional on month and year fixed effects.

32

.03

.05

.07

Aver

age

mar

gina

l cos

t

-3.5

-3-2

.5-2

-1.5

-1Av

g da

ily k

Wh

(est

'd c

oeff)

Jan Jul DecCalendar month

Treatment effect Average marginal cost

Figure 5: Treatment effects by month

Notes: Month-specific treatment effects, from OLS regression with customer and month-year fixedeffects.

33

27 4 18 1122 2 20 623 13 17 16Postpaid PV0

100%

200%

300%

400%

Prep

aid

retu

rns

rela

tive

to p

ostp

aid

Dom

estic

Life

line

High

kW

hLo

w kW

h

High

pro

p va

lLo

w pr

op v

al

Usua

lly o

n tim

eUs

ually

late

No d

ebts

Out

stan

ding

deb

ts

Neve

r disc

onne

cted

Disc

onne

cted

Figure 6: Heterogeneity in relative returns to prepaid metering

Notes: Returns to prepaid metering relative to postpaid metering per customer per month. Separateeffects are calculated for each customer characteristic (each pair of bars splits the population), andstandard errors are bootstrapped. Heterogeneity variables are as described in Table 3. The presentvalue of the monthly returns to the utility for each customer type on postpaid metering are printedbelow the bars.

34

1516

1718

19Pr

esen

t val

ue

0 .05 .1 .15 .2 .25Annual interest rate

Postpaid Prepaid

-50

050

Pres

ent v

alue

0 1 2 3X times marginal costs (USD)

Postpaid Prepaid

Figure 7: Returns, under alternative cost assumptions

Notes: The present value per month, by meter type. The top figure shows results as interest ratesvary and the bottom figure shows results by multiples of the observed average marginal costs ofsupply.

35

Table 1: Sample statistics

Number Mean Median SD Min MaxTotal obs. 211,190 - - - - -Total cust 4,246 - - - - -Obs. per cust - 49.74 54 8.989 1 54Prepaid obs. per cust - 15.25 16 7.596 0 22

Pre-project customer-level statisticsDaily kWh - 16.18 15 8.20 0 101Amount owed per mo. - 52 49 35 0 404Days to pay - 97.54 67 131.08 -969 1,381Share paid late - 0.54 0.58 0.33 0 1Outstanding debts - 0.26 - - 0 1Ever disconnected - 0.21 - - 0 1Lifeline tariff - 0.31 - - 0 1Property value - 29,856 26,662 14,115 10,704 156,908

Notes: Summary statistics for the sample. Pre-project customer level statistics are calculated for allobservations before November 2014 at the customer level. Days to pay is calculated as the numberof days between when consumption occurs and payment is received by the utility. Share paid lateis the share of months in which the customer paid past the billing due date. Outstanding debtsis an indicator for whether the customer had multiple unpaid bills at the end of the panel. Everdisconnected equals one if payment was ever enforced through disconnection. Lifeline is the shareof customers that had ever received the Lifeline tariff prior to the program. Property values areassessed values for tax purposes. All monetary units are in 2014 USD.

36

Table 2: Average daily kWh

OLS IV(1) (2) (3) (4)

Panel A: Avg daily kWhPrepaid -2.135*** -2.173*** -1.964*** -1.889***

(0.110) (0.137) (0.110) (0.199)R

2 0.150 0.153 0.149 0.153N 207,930 207,930 207,929 207,929N customers 4,246 4,246 4,245 4,245Fixed effects month, year month-year month, year month-year

Panel B: Log avg daily kWhPrepaid -0.135*** -0.143*** -0.119*** -0.152***

(0.010) (0.012) (0.010) (0.016)R

2 0.086 0.089 0.086 0.089N 207,010 207,010 207,009 207,009N customers 4,246 4,246 4,245 4,245Fixed effects month, year month-year month, year month-year

Notes: The table shows the effect of prepaid metering on average daily kWh per customer per monthin levels (Panel A) and logs (Panel B). Columns 1 and 2 report OLS coefficients on an indicatorfor prepaid metering, columns 3 and 4 instrument for the switch date with assignment to prepaidmetering. See the main text for further description of the instrument. All specifications excludethe switch month and include customer fixed effects and a time-varying tariff control and clusterstandard errors at the customer level. Odd numbered columns include calendar month and yearfixed effects; even numbered columns include month-year fixed effects.

37

Table 3: Heterogeneous treatment effects

OLS IV

N Mean Total effects Difference Total effects Difference(1) (2) (3) (4)

Domestic 2,907 17.71 -2.169*** 1.260*** -1.930*** 1.211***(0.137) (0.203) (0.125) (0.255)

Lifeline 1,326 9.34 -0.909*** -0.720***(0.151) (0.222)

Above median kWh 2,132 20.05 -2.652*** 1.626*** -2.590*** 1.923***(0.170) (0.208) (0.154) (0.208)

Below median kWh 2,101 10.06 -1.026*** -0.667***(0.121) (0.140)

High prop value 1,482 17.54 -1.837*** -0.494** -1.674*** -0.460**(0.195) (0.235) (0.182) (0.228)

Low prop value 2,751 13.77 -2.331*** -2.133***(0.132) (0.138)

Usually on time 2,103 14.03 -1.752*** -0.818*** -1.751*** -0.508**(0.128) (0.221) (0.128) (0.216)

Usually late 2,130 16.14 -2.570*** -2.259***(0.181) (0.174)

No debts 3,131 15.12 -2.132*** -0.092 -2.092*** 0.371(0.111) (0.286) (0.114) (0.311)

Outstanding debts 1,102 15.01 -2.225*** -1.721***(0.264) (0.289)

Never disconnected 3,363 14.84 -2.205*** 0.121 -2.122*** 0.575*(0.110) (0.334) (0.111) (0.342)

Ever disconnected 870 16.07 -2.084*** -1.547***(0.316) (0.324)

Notes: Effects of the prepaid meter on average daily kWh by sub-group. Each coefficient is froma separate regression. Columns 1 and 2 report OLS coefficients and Columns 3 and 4 report IVcoefficients. Specifications include separate month and year fixed effects interacted with the hetero-geneity variable to allow for differential seasonal time trends by characteristics. All characteristicsare defined by the pre-project period as follows: Lifeline equals one for customers primarily onlifeline tariff, Low prop value equals one for customers with a 2012 ZAR property value below300,000, Usually late equals one for customers who paid above the median share of their monthlybills past the due date, Outstanding debts indicates that the customer had multiple unpaid billsat the end of the panel. Ever disconnected equals one if the customer was ever disconnected onpostpaid metering. 38

Table 4: Revenue-related outcomes

Amount owed Avg marg cost Days to pay(1) (2) (3)

Panel A: OLSPrepaid -6.909*** -2.920*** -74.641***

(0.548) (0.248) (4.001)

Panel B: IVPrepaid -6.787*** -2.805*** -84.645***

(0.788) (0.361) (4.946)

N obs 207,930 207,930 200,141N customers 4,246 4,246 4,234Month-year FE x x x

Notes: The table shows the effect of prepaid metering on the amount owed for consumption (column1), the monthly kWh supply cost to the City (column 2), and the days between consumption andpayment (column 3).

39

Table 5: Present value calculation (i = 0.08)

Postpaid Prepaid

Revenue estimateRevenue (pqm) 51.13 45.20

(0.26) (0.26)Includes:- Vendor commission 0 0.002/kWh

Payment timing estimateMonths since consumption (sm) 2.71 -0.11

(0.05) (0.05)

Cost estimateCost (cqm + bm) 33.68 27.01

(0.13) (0.13)Includes:- Meter reading 0.91 0- Billing 0.85 0- Losses (percent) 11.25 11.25- Disconnections 90 0

Present value 16.46 18.11(0.31) (0.34)

Notes: Present value in USD2014 by metering type for the City of Cape Town. Predicted meansand standard errors (delta method) are reported for each estimate, together with a list of theadministrative inputs to the calculations. The bottom panel presents the present value calculationper month at an annual interest rate of 8 percent, with bootstrapped standard errors.

40

A.1 Appendix tables and figures

41

Figure A.1: Prepaid electricity receipts - Lifeline customer

42

0.05

.1.15

USD/kWh

0 50 150 350 600kWh

Lifeline   Domestic

Jul  2011  -­  Jun  2012  (in  2014  USD)

0.05

.1.15

USD/kWh

0 50 150 350 600kWh

Lifeline   Domestic

Jul  2012  -­  Jun  2013  (in  2014  USD)

0.05

.1.15

.2USD/kWh

0 25 60 350 600kWh

Lifeline  1 Lifeline  2Domestic

Jul  2013  -­  Jun  2014  (in  2014  USD)

0.05

.1.15

.2USD/kWh

0 25 60 350 600kWh

Lifeline  1 Lifeline  2Domestic

Jul  2015  -­  Jun  2016  (in  2014  USD)

Figure A.2: Electricity tariffs

43

Civ

ic C

entr

e1

2 H

ertz

og

Bo

ule

vard

80

01

PO B

ox

65

5 C

ape

Tow

n 8

00

0V

AT

Reg

istr

atio

n n

um

ber

45

00

19

34

97

022110588401082014

Tax

invo

ice

nu

mb

er

Cu

sto

mer

VA

T re

gis

trat

ion

nu

mb

er

Acc

ou

nt

nu

mb

er

Dis

trib

utio

n c

od

e

Bu

sin

ess

par

tner

nu

mb

er

Co

py

ta

x i

nv

oic

e

Tel:

08

6 0

10

30

89

- F

ax:

08

6 0

10

30

90

Tel:

ove

rsea

s cl

ien

ts +

27

21

40

1 4

70

1E-

mai

l: ac

cou

nts

@ca

pet

ow

n.g

ov.

zaC

orr

esp

on

den

ce :

Direc

tor

Rev

enu

ePO

Bo

x 6

55

Cap

e To

wn

80

00

Web

ad

dre

ss:

ww

w.c

apet

ow

n.g

ov.

za

Acc

ou

nt

sum

ma

ry a

s a

t 0

1/0

8/2

01

4D

ue

da

te 2

6/0

8/2

01

4

At

/ E

rf

Prev

iou

s ac

cou

nt

bal

ance

39

64

.99

Less

pay

men

ts (

31

/08

/20

14

)Th

ank

Yo

u3

96

4.9

9-

(a)

0.0

0

Arr

ang

emen

t o

n a

rrea

rs o

f7

41

7.6

9

Cu

rren

t in

stal

men

t2

30

0.0

02

30

0.0

0

Cu

rre

nt

am

ou

nt

du

e (

b)

Pa

ya

ble

by

26

/08

/20

14

42

98

.16

To

tal

(a)

+ (

b)

42

98

.16

Tota

l (a)

+ (

b)

abo

ve4

29

8.1

6

Arr

ang

emen

t o

n a

rrea

rs o

f5

11

7.6

9

To

tal

lia

bil

ity

94

15

.85

Cu

rren

t ch

arg

es

tota

llin

g R

4298.1

6 w

ill b

e d

eb

ited

fro

m y

ou

r b

an

k a

cco

un

t.Ple

ase

co

mm

en

t o

n t

he 2

016/1

7 d

raft

bu

dg

et

du

rin

g A

pri

l. D

eta

ils

availa

ble

fro

m A

pri

l 4

at

sub

cou

nci

l o

ffic

es

an

d l

ibra

ries.

Fo

rm

ore

in

form

ati

on

go

to

htt

p:/

/ww

w.c

ap

eto

wn

.go

v.z

a/e

n/B

ud

get/

Pag

es/

Bu

dg

et-

2016-1

7.a

spx.

Ple

ase

no

te:

1.

2.

3.

4.

Paym

ent

options

(a)C

heq

ues

: (i)

Mad

e pay

able

to t

he

City

of

Cap

e To

wn. (ii

) Li

mit o

f R500 0

00,0

0. (ii

i) Po

st-d

ated

cheq

ues

are

not

acce

pta

ble

.(b

)Deb

it o

rder

s: C

all 0

860 1

03 0

89 o

r vi

sit

a C

ust

om

er S

ervi

ce C

entr

e. (c)

Inte

rnet

pay

men

ts: V

isit E

asyp

ay.c

o.z

a or

pay

City.

co.z

a.(d

)Ele

ctro

nic

pay

men

ts (EF

T): Se

lect

the

City

of

Cap

e To

wn a

s a

ban

k-lis

ted b

enef

icia

ry o

n y

our

ban

k's

web

site

. U

se o

nly

your

nin

e-dig

itm

unic

ipal

acc

ount

num

ber

as

refe

rence

.(e)

Direc

t dep

osi

t at

ABSA

Ban

k: P

leas

e in

sert

your

acco

unt

no. 221105884 o

n d

eposi

t sl

ip.

(f)D

ebit/c

redit c

ard: O

nly

at

City

vendors

ref

lect

ed b

elow

. Ban

k ch

arges

will

be

levi

ed o

n y

our

acco

unt

in r

espec

t of

single

as

wel

l as

multip

le p

aym

ents

in e

xces

s of

R5 0

00,0

0.

Inte

rest

will

be

char

ged

on a

ll am

ounts

still

outs

tandin

g a

fter

the

due

dat

e.Y

ou m

ay n

ot

withhold

pay

men

t, e

ven if

you h

ave

subm

itte

d a

quer

y to

the

City

conce

rnin

g t

his

acc

ount.

Failu

re t

o p

ay c

ould

res

ult in

your

wat

er a

nd/o

r el

ectr

icity

supply

bei

ng d

isco

nnec

ted/r

estr

icte

d. Im

med

iate

rec

onnec

tion o

f th

esu

pply

aft

er p

aym

ent

cannot

be

guar

ante

ed. A

dis

connec

tion f

ee w

ill b

e ch

arged

and y

our

dep

osi

t am

ount

mig

ht

be

incr

ease

d.

Pay

poin

ts:

City

of C

ape

Tow

n ca

sh o

ffice

s or

the

vend

ors

belo

w;

>>

>>

>

Acc

ou

nt

nu

mb

er

Am

ou

nt

du

e if n

ot

pai

d in

cas

h

Am

ou

nt

du

e if p

aid

in c

ash

Ro

un

ded

do

wn

am

ou

nt

carr

ied

fo

rwar

d t

o n

ext

invo

ice

42

98

.16

42

98

.10

0.0

6

Page

1 o

f 4

000009

Late

st a

cco

un

t -

see

ove

rlea

f1

99

8.1

6

isu

inv_

20

16

04

01

_23

40

03

.mu

ln0

00

00

5 /

00

00

09

>>

>>

>

Page

2 o

f 4

Acc

ou

nt

de

tail

s a

s a

t 0

1/0

8/2

01

4A

cco

un

t n

um

be

r

PR

OP

ER

TY

RA

TE

S (

Pe

rio

d 0

2/0

7/2

01

4 t

o 0

1/0

8/2

01

4 )

31

Da

ys

At

# T

ota

l val

ue

Fro

m 0

2/0

7/2

01

4 :

R 1

70

00

00

.00

@ 0

.00

62

54

0 ÷

36

5 x

31

90

2.9

7#

Sta

tuto

ry r

ebat

e cr

edit

Fro

m 0

2/0

7/2

01

4 :

R 1

50

00

.00

@ 0

.00

62

54

0 ÷

36

5 x

31

7.9

7-

# A

dd

itio

nal

reb

ate

cred

itFr

om

02

/07

/20

14

: R

18

50

00

.00

@ 0

.00

62

54

0 ÷

36

5 x

31

98

.26

-7

96

.74

ELE

CT

RIC

ITY

( P

eri

od

27

/05

/20

14

to

24

/07

/20

14

-

59

Da

ys

) (A

ctu

al

rea

din

g)

At

Me

ter

no

: /

Co

nsu

mp

tio

n 1

18

5.0

00

kW

h /

Da

ily

av

era

ge

20

.08

5 k

Wh

*C

on

sum

ptio

n c

har

ge:

do

mes

tic

(1)

17

2.6

03

0 k

Wh

@ R

1.2

50

0 (

2)

23

0.1

37

0 k

Wh

@ R

1.2

50

0(3

) 2

87

.67

10

kW

h @

R 1

.25

00

(4

) 1

2.5

55

1 k

Wh

@ R

1.5

20

08

82

.09

*C

on

sum

ptio

n c

har

ge:

do

mes

tic

Fro

m 0

1/0

7/2

01

4 :

(1

) 1

18

.35

60

kW

h @

R 1

.34

76

(2

) 1

57

.80

80

kW

h @

R 1

.34

76

(3)

19

7.2

61

0 k

Wh

@ R

1.3

47

6 (

4)

8.6

08

9 k

Wh

@ R

1.6

38

76

52

.10

Rev

ersa

l of

estim

ated

co

nsu

mp

tio

n (

53

7.1

23

kW

h )

67

1.4

0-

86

2.7

9

WA

TE

R (

Pe

rio

d 2

8/0

6/2

01

4 t

o 2

4/0

7/2

01

4 -

2

7 D

ay

s )

(Act

ua

l re

ad

ing

)

At

Me

ter

no

: /

Co

nsu

mp

tio

n 1

1.0

00

kl

/ D

ail

y a

ve

rag

e 0

.40

7 k

l*

Co

nsu

mp

tio

n c

har

ge

(do

mes

tic)

Fro

m 2

8/0

6/2

01

4 :

(1

) 0

.59

20

kl f

ree

(2)

0.4

44

0 k

l @ R

7.6

00

0(3

) 0

.18

62

kl @

R 1

1.6

10

0Fr

om

01

/07

/20

14

: (

1)

4.7

34

0 k

l fre

e (2

) 3

.55

10

kl @

R 8

.75

00

(3)

1.4

92

8 k

l @ R

12

.54

00

55

.32

55

.32

RE

FUS

E (

Pe

rio

d 0

2/0

7/2

01

4 t

o 0

1/0

8/2

01

4 )

31

Da

ys

At

*R

efu

se c

har

ge

( 1

X 1

40

L R

ECY

L X

1 R

emo

vals

)0

.00

*R

efu

se c

har

ge

( 1

X 2

40

lBIN

X 1

Rem

ova

ls )

95

.96

95

.96

SE

WE

RA

GE

( P

eri

od

28

/06

/20

14

to

24

/07

/20

14

-

27

Da

ys

) (A

ctu

al

rea

din

g)

At

*D

isp

osa

l ch

arg

eFr

om

28

/06

/20

14

: (

1)

0.4

14

0 k

l fre

e (2

) 0

.31

10

kl @

R 7

.20

00

(3)

0.1

30

6 k

l @ R

13

.56

00

Fro

m 0

1/0

7/2

01

4 :

(1

) 3

.31

40

kl f

ree

(2)

2.4

85

0 k

l @ R

8.2

50

0(3

) 1

.04

54

kl @

R 1

4.6

40

03

9.8

23

9.8

2

Ad

d 1

4%

VA

T o

n a

mo

un

ts m

ark

ed

wit

h *

ab

ov

e1

47

.53

0%

VA

T o

n a

mo

un

ts m

ark

ed

wit

h #

ab

ov

e

Cu

rre

nt

acc

ou

nt:

To

tal

du

e1

,99

8.1

6

Me

ter

de

tail

sP

rev

iou

s re

ad

ing

Ne

w r

ea

din

gU

nit

s u

sed

ELE

CTR

ICIT

Y36191.0

00kW

h(A

ctu

al)

37376.0

00kW

h(A

ctu

al)

1185.0

00kW

h

WA

TER

242.0

00kl(

Act

ual)

253.0

00kl(

Act

ual)

11.0

00kl

Figu

reA

.3:

Sam

ple

cons

olid

ated

bill

44

.04

.05

.06

.07

.08

.09

Aver

age

mar

gina

l cos

t (U

SD)

2012m1 2013m1 2014m1 2015m1 2016m1

Figure A.4: City of Cape Town average marginal costs

Notes: Average marginal cost of electricity supply per month between 2012 and 2016, in USD2014.

45

Figure A.5: Randomization groups

Notes: Map of Cape Town. The polygons correspond to the 27 randomization groups. The 13groups that make up Mitchell’s Plain are clustered in the lower center of the map. Each polygoncontains between 150 and 200 customers.

46

010

0020

0030

0040

00N

o. o

f out

ages

2011m9 2012m7 2013m5 2014m3 2015m1 2015m11date

Load-shedding outages All other outages

01

23

45

Hou

rs

2011m9 2012m7 2013m5 2014m3 2015m1 2015m11Year-Month

All other outages Load-shedding outages

Figure A.6: Load-shedding events in Mitchell’s Plain

Notes: Description of loadshedding events and other outages in Mitchells Plain. The left figureshows the total number of separate incidents per month and the right figure shows the total hoursper month of each type of outage event.

-2-1

01

2-2

-10

12

-2-1

01

2-2

-10

12

-2-1

01

2

2012m1 2013m1 2014m1 2015m12012m1 2013m1 2014m1 2015m12012m1 2013m1 2014m1 2015m1

2012m1 2013m1 2014m1 2015m12012m1 2013m1 2014m1 2015m12012m1 2013m1 2014m1 2015m1

1 2 3 4 5 6

7 8 9 10 11 12

13 14 15 16 17 18

19 20 21 22 23 24

25 26 27

Residual

Figure A.7: Pre-program kWh residuals

Notes: Residuals from a regression of pre-program average daily kWh on customer and month-yearfixed effects, by randomization group.

47

1314

1516

1718

Avg

daily

kW

h

2012m1 2013m1 2014m1 2015m1 2016m1

Comparison group Project sample

Figure A.8: Average daily kWh, project versus comparison customers

Notes: Monthly mean consumption for project and comparison customers. The comparison group isa sample of postpaid customers, matched on property value. The vertical line in late 2014 representsthe start of the meter replacement program.

48

Table A.1: Balance

Switch DateGroup Assigned Actual Switched

Daily kWh -0.058* 0.002 -0.002 0.514(0.032) (0.005) (0.003) (0.315)

Lifeline tariff 0.007*** 0.000 0.000 -0.041**(0.002) (0.000) (0.000) (0.019)

Property value -21.666 64.269*** 16.454*** -3.9e+03***(55.855) (8.242) (4.537) (540.824)

Notes: Correlations between program administrative variables and pre-program customer charac-teristics, at the customer level (N=4212). Column 3 is conditional on switching to a prepaid meter(N=3213).

Table A.2: Average daily kWh - Mitchells Plain only

OLS IV(1) (2) (3) (4)

Panel A: Avg daily kWhPrepaid -2.002*** -2.127*** -1.887*** -2.224***

(0.126) (0.198) (0.117) (0.383)R

2 0.189 0.194 0.189 0.194N 112,050 112,050 112,050 112,050N customers 2,251 2,251 2,251 2,251Fixed effects month, year month-year month, year month-year

Panel B: Log avg daily kWhPrepaid -0.117*** -0.132*** -0.106*** -0.159***

(0.013) (0.023) (0.011) (0.038)R

2 0.108 0.112 0.108 0.111N 111,725 111,725 111,725 111,725N customers 2,251 2,251 2,251 2,251Fixed effects month, year month-year month, year month-year

Notes: Consumption results for the Mitchells Plain sample only. Otherwise, details are as in Table2.

49

Tabl

eA

.3:

Cor

rela

tion

betw

een

hete

roge

neity

varia

bles

Life

line

Low

kWh

Low

prop

valu

eU

sual

lyla

teU

npai

dbi

llsD

iscon

nect

edLi

felin

e1

Low

kWh

0.61

1⇤⇤⇤

1Lo

wpr

opva

lue

0.21

9⇤⇤⇤

0.22

2⇤⇤⇤

1U

sual

lyla

te-0

.114

⇤⇤⇤

-0.1

41⇤⇤

⇤0.

0203

1U

npai

dbi

lls-0

.007

590.

0201

0.08

41⇤⇤

⇤0.

307⇤

⇤⇤1

Disc

onne

cted

-0.0

709⇤

⇤⇤-0

.037

9⇤-0

.040

1⇤⇤

0.35

3⇤⇤⇤

0.29

9⇤⇤⇤

1

Not

es:

Pair

wis

eco

rrel

atio

nsof

vari

able

sus

edin

the

hete

roge

neity

anal

ysis

,at

the

cust

omer

leve

l.A

llch

arac

teri

stic

sar

ede

fined

byth

epr

e-pr

ojec

tpe

riod

asfo

llow

s:Li

felin

eeq

uals

one

for

cust

omer

spr

imar

ilyon

lifel

ine

tari

ff,Lo

wkW

heq

uals

one

ifth

eav

erag

eda

ilykW

hm

easu

reis

belo

wth

em

edia

n,Lo

wpr

opva

lue

equa

lson

efo

rcu

stom

ers

wit

ha

2012

ZAR

prop

erty

valu

ebe

low

300,

000,

Out

stan

ding

debt

seq

uals

one

for

cust

omer

sw

ith

mul

tipl

eun

paid

bills

atth

een

dof

the

pane

l,U

sual

lyla

teeq

uals

one

for

cust

omer

sw

hopa

idab

ove

the

med

ian

shar

eof

thei

rm

onth

lybi

llspa

stth

edu

eda

te.

Dis

conn

ecte

deq

uals

one

for

cust

omer

sth

atw

ere

ever

disc

onne

cted

onpo

stpa

id.

50

Table A.4: Robustness checks (average daily kWh)

Base Group IV Mailing IV DD Postpaid DD Prepaid(1) (2) (3) (4) (5)

Panel A: OLSPrepaid -2.173*** – – -2.271*** -1.856***

(0.137) (0.117) (0.087)

Panel B: IVPrepaid -1.889*** -1.746*** -1.781*** -2.325*** -1.718***

(0.199) (0.430) (0.273) (0.172) (0.100)

N obs 207,930 207,929 207,929 271,137 2,004,004N customers 4,246 4,245 4,245 5,452 38,248Month-year FE x x x x x

Notes: Robustness to alternative specifications. The base result (column 1) corresponds to columns2 and 4 of Panel A in Table 2. Column 2 uses the group order as the instrument (equal to zero priorto the start of the program in November 2014). Column 3 uses the date of the mailing informingcustomers of the program as the instrument. Column 4 adds a comparison group of postpaidcustomers, not in the program (i.e. never switched), sampled based on property value. Column 5adds a comparison group of prepaid customers in the project areas. See text for further details.

51

Tabl

eA

.5:

Rob

ustn

ess

chec

ks(a

vera

geda

ilykW

h)

Bas

eSw

itche

rson

lyN

ode

btre

cove

ryTr

imm

edSm

ooth

edB

alan

ced

Bal

ance

d(1

)(2

)(3

)(4

)(5

)(6

)(7

)

Pane

lA:O

LSP

repa

id-2

.173

***

-2.2

41**

*-2

.192

***

-2.1

84**

*-2

.324

***

-2.1

35**

*-2

.066

***

(0.1

37)

(0.1

65)

(0.1

39)

(0.1

12)

(0.1

32)

(0.1

40)

(0.1

49)

Pane

lB:I

VP

repa

id-1

.889

***

-2.1

43**

*-1

.905

***

-2.1

58**

*-1

.987

***

-1.7

24**

*-2

.025

***

(0.1

99)

(0.1

83)

(0.2

04)

(0.1

72)

(0.1

94)

(0.2

09)

(0.2

11)

Nob

s20

7,93

016

0,08

320

4,82

820

3,73

820

7,05

220

9,07

517

0,67

8N

cust

omer

s4,

246

3,25

24,

245

4,23

44,

246

4,24

63,

213

Mon

th-y

ear

FEx

xx

xx

xx

Not

es:

Rob

ustn

ess

tosa

mpl

ean

dva

riab

leco

nstr

ucti

on.

The

base

resu

lt(c

olum

n1)

corr

espo

nds

toco

lum

ns2

and

4of

Pane

lAin

Tabl

e2.

Col

umn

2on

lyin

clud

escu

stom

ers

who

swit

ched

met

erty

pes.

Col

umn

3dr

ops

obse

rvat

ions

inw

hich

the

prep

aid

met

eris

used

tore

cove

rde

bts.

Col

umn

4tr

ims

the

top

1pe

rcen

tof

the

outc

ome

vari

able

.C

olum

n5

allo

ws

for

alo

nger

win

dow

over

whi

chbi

llsor

prep

aid

tran

sact

ions

are

aver

aged

.C

olum

n6

bala

nces

the

pane

lfor

allc

usto

mer

s,re

plac

ing

mis

sing

outc

omes

wit

hze

ros.

Col

umn

7ba

lanc

esth

epa

nelb

yex

clud

ing

cust

omer

sw

ith

mor

eth

anfo

urm

issi

ngm

onth

sof

data

.

52

Table A.6: Robustness checks (average daily kWh)

Base Load-shedding Tariff error Placebo test(1) (2) (3) (4)

Panel A: OLSPrepaid -2.173*** -2.200*** -2.490*** -0.153

(0.137) (0.138) (0.147) (0.121)

Panel B: IVPrepaid -1.889*** -1.886*** -1.626*** 0.213

(0.199) (0.199) (0.220) (0.193)

N obs 207,929 185,761 162,706 153,748N customers 4,245 4,243 3,316 4,213Month-year FE x x x x

Notes: Robustness to program implementation issues. The base result (column 1) corresponds tocolumns 2 and 4 of Panel A in Table 2. Column 2 controls for the average daily hours of regular andload-shedding outages in the month and is restricted to the Mitchell’s Plain sample of customers.Column 3 drops customers with tariff mistakes. Column 4 implements a placebo check that movesthe assigned switch date and actual switch date ahead by one year.

53

Table A.7: Heterogeneity in returns to prepaid metering

Average returns Relative returnsPostpaid Pre / Post

Domestic 22 1.11(0.01)

Lifeline 2 1.41(0.37)

Above median kWh 27 1.05(0.02)

Below median kWh 4 1.61(0.10)

High prop value 23 1.03(0.02)

Low prop value 13 1.18(0.03)

Usually on time 17 0.93(0.01)

Usually late 16 1.29(0.04)

No debts 20 0.90(0.01)

Outstanding debts 6 3.59(0.47)

Never disconnected 18 0.98(0.01)

Outstanding debts 11 1.93(0.14)

Notes: Returns to prepaid metering relative to postpaid metering, by customer characteristic. SeeFigure 6 for further detail.

54

A.2 Data and variables

This appendix details the data sources and how they are combined, and a detailed descriptionof the variables used in the analysis.

A.2.1 Data sources and dataset construction

Billing records The City of Cape Town maintains billing records for any property servedor taxed by the municipality. As discussed in the main text, most households receive aconsolidated bill for all taxes and services every 25-35 days, with billing dates that varyacross customers. We create a billing panel that sequences bills by meter reading date. Theresulting panel contains both overlapping billing periods and gaps between billing periods.Overlapping billing periods are most commonly due to estimated meter readings (10.3 per-cent of bills in the raw data).26 Once an actual reading is collected, the estimated readingsare reversed and the customer is billed for the difference between the estimated and actualreadings during the estimated months. Actual readings are used to replace estimated read-ings in the data, by assigning the actual consumption estimated billing periods assumingequal consumption on each estimated day. Gaps between bills are less common (2.1 percentof bills in the raw data). Gaps and bills with zero recorded consumption are dealt withsimilarly in the cleaning process. We allow for two alternative assumptions: (1) averageover gaps of up to 30 days (including gaps associated with zero consumption bills), workingbackward from the date of the next non-missing (non-zero) bill, or (2) average over gaps ofup to 365 days using the same process. (1) is our main outcome measure, and (2) is used ina robustness check. All gaps longer than 365 days are dropped (N=102).

Prepaid vending records The prepaid vending system records each transaction and themeter with which it is associated. The meters themselves do not communicate with the grid,and as a result, we do not observe prepaid meter consumption directly. To construct monthlyoutcome measures comparable to those obtained through the billing records, we assume thatelectricity is consumed at a constant rate between purchases and that customers maintaina steady minimum balance (which may be zero) over time, i.e. there is no accumulation ofprepaid credit on the meter.

26Estimates are taken when a customer’s meter cannot be read, which usually occurs because it cannotphysically be accessed. Consumption is instead estimated based on past consumption patterns observed forthat customer. At most, three consecutive estimated readings are permitted by the system before an actualreading is obtained and used to “reverse” the estimated readings.

55

Customers purchase electricity frequently: the median frequency is every 3.3 days. Outof over 50,000 customer-month observations on prepaid metering, only 270 months are as-sociated with no prepaid purchases, corresponding to 147 unique accounts. Consequently,any more sophisticated latent demand model would only affect the assumed within-monthvariation in demand, which we cannot observe on either the prepaid or postpaid system.We impose analogous averaging assumptions to what is described above for the billing panelto address gaps between prepaid purchases of over a month. We allow for two alternativeassumptions: (1) average over gaps of up to 30 days, working forward from the last observedpurchase (i.e. assume entire transaction is consumed within 30 days), or (2) average overgaps of up to 365 days using the same process. Gaps of longer than 365 days are dropped.(1) is our main outcome measure, and (2) is used in robustness checks.

Project data The contractor maintained records of attempted and completed meter in-stallations, which we use to match postpaid and prepaid meters. Contractor records alsoinclude the date of meter installation, the meter serial number and the date that householdsreceived maildrops informing them of the project.

Sample construction and randomization used lists of targeted accounts provided by theDepartment of Electricity. We include all accounts that were on the lists in our analysis, withthe following exceptions. First, non-domestic customers are dropped. Second, customerswith 3-phase electricity meters were dropped. The contractor did not replace this type ofmeter. Finally, 13 meters in the randomization file that did not receive any bills betweenJanuary 2012 and November 2014 and were not in the contractor installation logs weredropped from the sample.

A.2.2 Variables

• Average daily kWh: We construct an average daily kWh variable at the customer-month level. As described above, our main variable averages over up to 30 days priorto the most recent meter reading or since the most recent prepaid purchase in thecase of months with no data. As a robustness check, we allow for a longer averagingwindow, of up to one year. We also use the total kWh consumed in the month in ourbenefit-cost analysis. We construct a binary indicator for above median kWh based onthe customer’s average consumption prior to November 2014.

• Amount owed: We apply the customer’s tariff to the constructed consumption measure,

56

calculating the kWh on each tariff block and the marginal price. This results in anamount owed associated with the calendar month of consumption.

• Days to pay: We construct a variable that describes the number of days between whena customer consumes electricity and when he or she pays for that electricity. Forprepaid observations, this is calculated as half of the average number of days betweentransactions, consistent with the assumption of a constant rate of consumption betweentransactions. For postpaid observations, we take the amount owed on the first bill inthe panel and use that as the starting balance that must be cleared. A bill is clearedwhen cumulative payments catch up with the cumulative amount owed. For customersthat receive a consolidated bill, accounting is similar, though debts must also be clearedbefore a payment is allocated toward electricity.27 The days to pay is transformed intoa months to pay variable for the benefit cost analysis. We also use this variable toconstruct late payment measures, which equal one if the bill was paid off after its duedate. A customer is categorized as usually late if over 58 percent (the median share)of bills before November 2014 are paid late.

• Average marginal cost: We obtain records of the average marginal cost paid eachmonth by the City of Cape Town to Eskom. This is calculated based on the time ofconsumption for all residential and commercial customers in the City.

• Non-payment: For bills that are not cleared by the end of the panel, we construct apayment probability variable based on observed payment probabilities associated withdebts of different ages in a longer panel for the same sample. This payment probabilityis set to zero for debts older than 3 years, as per South Africa’s Municipal Systems Act(i.e. debts older than 3 years are written off). For payments that we do not observe,we set the revenue measure in our benefit cost analysis equal to the amount owedtimes the payment probability. We use the customer’s average time to pay to replaceunobserved days to pay. We construct a measure of outstanding debts that equals oneif the customer has multiple unpaid bills at the end of the panel.

• Disconnections: Customers are charged for disconnections and reconnections associatedwith enforcing payment. We record the cost of a disconnection in the month that it

27The City of Cape Town assigns payments against the consolidated bill to debt first, followed by electricity,then other services. We therefore assume that the electricity amount owed is cleared once cumulativepayments catch up with the cumulative amount owed from past bills plus the current owed for electricityonly.

57

shows up on the customer’s bill. The disconnection costs to the City are factored intothe benefit cost analysis. We construct an indicator for whether the customer receivedany disconnections on their postpaid meter.

• Property value: We use the City of Cape Town’s 2012 general valuation of properties,which is the basis for property taxes, along with a geographic identifier to matchproperty values to electricity meters. Our binary measure of low property value uses athreshold of 300,000 ZAR, which is the cutoff for several social programs in the City.We assume low values for flats and for a small number of parcels with missing data.

• Administrative cost records: Other details included in the benefit cost analysis wereobtained from the City of Cape Town through personal communication with the Elec-tricity Department. These include the rate of technical and non-technical losses, andthe cost of preparing bills and reading meters.

58